OpenWiki synthesis

Type: agentic-system-analyses/types/agentic-system-synthesis.md

Bounded synthesis

OpenWiki produces and maintains linked Markdown knowledge from repository source or personal sources, then supplies retained pages to humans and agents. This account covers its shipped production, maintenance, admission and consumption paths at commit 7c469e59584bfac3f97f5c32a67604de46805999, with cutoff 2026-10-14. SRC-1 supplies inspected implementation and source-native design contracts, not execution traces or experimental results. No target, test, provider or evaluator was executed. External coding hosts, dependency implementations, served model parameters and live connector services are outside the frozen evidence. The whole-system boundary describes functional scope, not exhaustive inspection. See the boundary and coverage and Source register.

Request, research and assigned work

The main architectural division is between semantic authoring and deterministic lifecycle control. For native repository init or update, code begins or resumes a durable run, returns a strict no-op when applicable, and otherwise launches a planner. The model researches source and proposes a page taxonomy; code admits the canonical plan and allocates pending page jobs. Fresh page agents receive selected assignments and source tools. OpenWiki owns queue transitions; narration cannot declare completion. The external-host alternative uses the same lifecycle but delegates research, model session and physical page writes to the coding host. See RT-RTE-native-generation and RT-RTE-host-generation.

The native workflow is representative because the shipped runner connects planning, page delegation, rejection, recovery and delivery. Each page assignment includes purpose and instructions, existing-Claim count and Claims needing attention; the full current set is available by explicit inspection. Only quickstart receives the complete planned page map. Native model-facing writes are limited to the assigned page, and model-facing task delegation is removed. Parallel job ownership is process-local, not a cross-process lease; quickstart is deferred to a final single-slot pass. These are source-level controls, not demonstrated model compliance or operating-system isolation. See RT-OBJ-page-guidance, RT-CLM-native-confinement and the workflow account.

Fresh initial graphs do not mean all predecessor research is erased. Native planner/page backends share a durable home-directory conversation-history mount. Model-tool write denial there does not prevent direct middleware backend writes; backend reads afford later access. Actual offload contents, summarization policy, automatic replay and cross-worker use remain unresolved. Chat separately uses persistent SQLite checkpoints and same-thread continuation. Pruning old database snapshots proves neither removal of old turns inside surviving state nor deletion of history files. See MEM-OBJ-conversation-history, MEM-RTE-history-offload and MEM-RTE-chat-continuation.

Evidence debt, admission and delivery

Repository pages retain structured Claims: natural-language factual statements paired with code-resolved resource identities and version tokens. On a later update, comparisons identify stale or unresolved evidence and require affected-page work. An issue requires an explicit confirmation, update or retraction; omitted issue-free Claims survive automatically. This is an evidence-debt mechanism rather than a timestamp-only refresh. It checks the current identity of selected evidence, not whether that evidence entails the statement or covers every semantic dependency. See MEM-OBJ-claim, MEM-OBJ-evidence, EPI-RTE-evidence-preflight and EPI-RTE-claim-decisions.

Submission repairs/validates frontmatter, reconciles sparse decisions, rechecks evidence and persists Claims before recording page completion. Invalid input returns a corrective tool response. Ordinary unsubmitted work can receive one fresh retry from restored page/Claim snapshots, then be skipped and restored; rate limits bypass immediate retry, and fatal submission errors prevent finish. Snapshot recovery does not roll back shared history or raw external-host writes. Finish refuses pending jobs, checks skipped snapshots, finalizes artifacts and manifest coverage, and removes resumable state last. A complete finish may include skipped restored pages and interrupted update metadata; it does not mean every planned page was freshly regenerated. See RT-CLM-page-durability and the workflow account.

The resulting verification event records successful nonempty, durable reconciliation, not independent semantic acceptance. Source-to-proposition judgment and agreement between prose and Claims remain model/host responsibilities under shipped guidance. No produced, semantically accepted ampliative claim is evidenced by this analysis. Precise resource references remain in sidecars, while page-level provenance collapses line ranges into whole-file references. Removing an OpenWiki verification event does not remove prose, and foreign verification events are preserved without certification. Pending-page exclusions also do not guarantee untouched Markdown: projection can remove another pending page's stamp. See EPI-CLM-grounded-claims, EPI-RTE-claims-persistence, EPI-RTE-source-projection, EPI-RTE-verification-projection and report verification limits.

Local retention and Git adoption have different owners. The inspected scheduled workflow can propose completed partial progress in a PR even after CLI failure, while deleting transient queue state. Actual merging and deployed approval policy are unknown. Personal ingestion uses connector acquisition followed by model filesystem edits, not the repository Claim-submission gate; connector-derived facts are explicitly outside the Grounded Claims contract. MCP read-only eligibility trusts configuration or server descriptors, not an enforced guarantee of remote nonmutation. Direct translation checks nonempty/nontruncated output, not semantic fidelity. See RT-RTE-maintenance-admission, RT-RTE-personal-ingestion, RT-CLM-mcp-policy and RT-RTE-page-translation.

Later consumption and assessment

Coding-host retrieval progressively exposes retained knowledge. A request triggers current-Markdown reload and lexical section ranking; the host then requests complete selected sections. Search excerpts and result counts are bounded, but complete section reads have no inspected total token budget. Retrieval returns prose, not generated answers or sidecar proof packages. The host skill says to treat it as context rather than instructions and verify important details against current source. Deprecated pages are suppressed in search but remain accessible by direct section read. Thus relevance, current bytes and withdrawal at one consumer do not establish truth, freshness or universal withdrawal. See RT-RTE-wiki-retrieval, its memory annotation and RT-BAP-retrieved-sections.

For repeated repository documentation, the supported contribution is durable progress plus targeted evidence-debt repair and selective later context supply. For source-grounded answering, reliability still depends on semantic authoring and downstream source verification. For personal knowledge, import warrant and source-specific synchronization remain separately unresolved. Human visualization and benchmark drivers are real later-consumer wiring, not evidence that anyone used the knowledge successfully. DeepSWE can copy generated or compatible cached artifacts into a task workspace; this does not prove current-package regeneration on every run. See MEM-RTE-visualizer-delivery and MEM-RTE-evaluation-transfer.

The strongest supported trace-fed contribution is an automatic path from selected external operational traces to retained runtime documentation available to later agents. Sampling favors anomalies; compact records retain structure, metrics and errors, not full inputs/outputs. This supports a wired trace-fed document pathway, including learning authority at its update consumer, without establishing improved future capacity. Requested patterns, implications and incremental refinement do not establish new-claim synthesis or actual document refinement as curation operations. See MEM-OBJ-trace, MEM-RTE-trace-documentation and the profile verification.

Learning as improved capacity for future action remains unestablished: there is no demonstrated comparison linking retained criticism to greater later capacity. Theory-builder conditions remain independently unestablished in the assessed routes: condition 1 requires formulated theory content, not merely a factual Claim; condition 2 requires decisions dependent on that theory's content, not delivery alone; condition 3 requires stated content-directed criticism with blame and changed reliance or revision; condition 4 requires its retained result to shape another round. The records establish Claim maintenance and requested trace/source criticism, not those four findings. Criticism of a consumed theory improving future capacity is likewise unestablished. Reflection is not established by documenting external runtimes; self-representation and both causal directions would need evidence. Autonomous theory building is not established by automatic queue/admission control. These are separate uncertainties, not a combined negative verdict. See RT-RTE-native-generation, MEM-RTE-trace-documentation and the epistemic bounded conclusion.

At this boundary, exercised self-improvement is also unestablished. The trace pathway may change knowledge relevant to future action, but external-runtime documentation is not automatically a change to OpenWiki's own behavior-determining organization. No supplied operation closes the link from such a change to subsequent OpenWiki behavior causally dependent on it. Neither successful improvement nor a fully established dispositional self-improvement pathway follows. The served model identity also is not an immutable weight pin, and local parameter updating is not established by the inspected invocation paths. See MEM-RTE-trace-documentation and RT-CMP-chat-model.

LEDGER provides a distinct semantic evaluator: model judgments compare extracted assertions with bounded current/historical evidence, then feed metrics, caches and later evaluation watch state. The inspected runner does not feed those verdicts into the next product update as an admission or revision gate. Its existence prevents claiming that OpenWiki ships no semantic evaluator, but supplies no measured accuracy or causal benefit. See EPI-RTE-ledger-judgment, EPI-RTE-ledger-reporting and MEM-RTE-ledger-evaluation.

The assessment would change with proposition-level input/output and evidence-consuming acceptance traces; production dependency inspection resolving history payload and replay; downstream traces linking retained sections to decisions; and controlled comparisons of capacity or outcomes at the actual intervention level. A semantic admission gate or a projector that receives pending exclusions would change the respective source-level guarantees. No direct source reads were made during synthesis; supporting passages remain in the cited accepted records.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Pending exclusions do not propagate to verification projection. An absent eligibility event can remove another pending page's OpenWiki stamp. This carries the report-verification limit; no current unresolved conflict or profile-verification limit remains.	SRC-1; RT-CLM-page-durability, EPI-RTE-verification-projection, workflow account, report verification limits	Shared submission, Claim-session exclusions and all-page metadata projection	Complete pending-page noninterference or uniform prevention of transient metadata writes; not per-submission durability ordering or final whole-run checks. No observed race, corruption or delivery failure follows.	Exclusion-aware projection implementation and bounded execution evidence for concurrent submissions.
Source-level wiring has no supplied execution or experiment results.	SRC-1; RT-RTE-native-generation, RT-RTE-host-generation, MEM-RTE-evaluation-transfer	Frozen shipped generation, publication and downstream-transfer code	End-to-end reliability, actual host uptake, compliance, task benefit or causal component effects	Inspectable executions with artifact lineage and outcomes; independently varied treatments for component attribution.
Fresh graphs, checkpoint cleanup and history infrastructure leave model-visible retained content unresolved.	SRC-1; MEM-OBJ-chat-checkpoint, MEM-RTE-chat-continuation, MEM-OBJ-conversation-history, MEM-RTE-history-offload	OpenWiki mounts, worker builders, SQLite selection/pruning and unexecuted offload contracts	Complete conversation isolation, history rollback/expiry, exact compaction, automatic replay or cross-worker reuse	Frozen dependency production/selection implementation and traces of actual retained payloads and later delivery.
Evidence identity and durability do not establish semantic warrant or complete factual coverage.	SRC-1; EPI-RTE-evidence-preflight, EPI-RTE-claim-decisions, EPI-RTE-source-projection, EPI-CLM-grounded-claims	Selected repository resources, sparse decisions, persistence and whole-file provenance	Statement entailment, prose/Claim agreement, accepted ampliative knowledge or semantic completeness; complete resolver relocation fidelity remains unassessed	Proposition-level source comparisons, acceptance decisions with criterion/use/scope, dependency coverage and resolver inspection.
Withdrawal and context budgets are consumer-specific.	SRC-1; RT-RTE-wiki-retrieval, MEM-RTE-claims-reconciliation, EPI-RTE-verification-projection	Search filtering, direct reads, current Claim membership and producer metadata	Universal withdrawal of prose/copies or guaranteed small full-section delivery	Shared status/reliance checks across consumers, bounded read budgets and traces of downstream withdrawal.
External hosts, providers and connector services remain outside the frozen implementation.	SRC-1; RT-CMP-chat-model, RT-RTE-host-generation, RT-CLM-mcp-policy	Client configuration, protocol and tool eligibility	Immutable served weights, provider update behavior, deployed host confinement/approval or remote nonmutation	Pinned external implementations, deployment policies and inspectable service/host operation.
Trace-fed documentation supports an architectural contribution, not improved capacity or theory criticism.	SRC-1; MEM-OBJ-trace, MEM-RTE-trace-documentation, profile verification	Anomaly-selected external records, compact transformation, requested authoring and later wiki access	Fleet-frequency estimates, actual new-claim synthesis/refinement, learning benefit, theory-builder conditions, reflection, autonomous theory building or exercised self-improvement	Actual authored propositions and criticism/revision lineage; retained-result consumers; self-representation links and objective-relative subsequent operation/comparison evidence.
Included alternatives remain partially inspected rather than absent.	SRC-1; MEM-OBJ-connector-data, MEM-RTE-visualizer-delivery, RT-RTE-maintenance-admission, MEM-RTE-evaluation-transfer	Shared connector I/O, graph construction, one scheduled driver and benchmark transfer	Uniform source synchronization/expiry, complete UI/export fidelity, scheduling equivalence, all initial admission/provenance controls or complete memory inventory	Source-specific transformations/cursors, browser/export and platform implementations, cache backing/trust and independent-entry provenance.
Semantic benchmark evaluation is separate from product revision, with incomplete evaluator coverage.	SRC-1; EPI-RTE-ledger-judgment, EPI-RTE-ledger-reporting, MEM-RTE-ledger-evaluation	Bounded source judgment, verdict reporting/cache/carry and checkpoint caller	Calibrated evaluator accuracy, complete extraction/selection/forgetting coverage, optional output-persistence classification or score-driven product improvement	Evaluator dependency and persistence inspection, calibrated results and an evidenced production feedback/consumer route.